A small team of six engineers completed a project in 76 days that was originally scoped for 30 developers over 12 to 18 months. This was not a hypothetical scenario. It happened when an Amazon Bedrock team stopped treating AI as a coding shortcut and started treating it as the foundation of how they work. The team shipped more production code in five months than in the previous ten years.
The gap between teams like this and the rest of the industry is widening quickly. AI coding agents have dramatically increased the rate at which software gets written, but not the rate at which it reaches customers. Commits are surging and CI/CD pipelines are busier than ever. Yet features shipped to production have not kept the same pace. The bottleneck is not the agent's ability to generate output. It is the agent's access to the knowledge it needs to make good decisions, and the team's willingness to restructure work around that reality.
Teams that have figured this out are called "frontier teams." They exist across industries and company sizes and share a common discipline: they treat AI adoption as an engineering investment, not a tool rollout. Any engineering team can become a frontier team, according to Swami Sivasubramanian, Vice President for Agentic AI at AWS.
Three paths to AI-native development at Amazon
AI-native software development treats AI as the foundation of how software is built, with increasingly capable agents directed by human experts. At Amazon, the primary drivers for AI in development were to reduce time developers spent on non-coding tasks such as documentation, coordination, and operations, retire technical debt, and minimize coding inconsistencies across thousands of small "two-pizza" teams. Experiments across hundreds of engineering teams identified three paths.
The pathfinder initiative was a controlled experiment. Six senior engineers received a single mandate: rebuild the Amazon Bedrock inference engine, originally estimated at 30 developers working 12 to 18 months. Instead of adding headcount, the team spent its first weeks redesigning workflows around AI. They shifted from discrete tasks to goal-driven outcomes, ran multiple agents in parallel, and set up systems for AI to work independently during off-hours. The project was delivered in 76 days. Individual developer productivity increased approximately 20x as measured by normalized commit velocity. Commits went from 2 per week to 40.
The structured sprint took a different approach. The Prime Video Financial Systems team ran a 10-day experiment. Six engineers worked in one room with zero context switching, no on-call duties, no other projects, and limited meetings. A senior engineer spent three weeks beforehand breaking complexity into well-scoped tasks with detailed requirements. Over 10 days, the team produced 556 commits against a baseline of 96 and reduced a 90-week project estimate to 24 weeks. That translates to nearly 6x throughput and 4x acceleration. They attributed the gain to three factors multiplying together: acceleration of low-judgment work (1.5x), higher focus on high-judgment work with no context-switching (1.5x), and instant access to agent-captured domain expertise (1.5x). Remove any one factor and the gains collapse.
The in-situ experiment studied more than 50 teams. The 25 teams that implemented both new tools and new practices outperformed those that simply added AI to existing workflows. Amazon Stores ran structured pilots with typical development teams working against their regular backlogs. The median productivity gain was 4.5x, with some teams reaching more than 10x improvement in normalized deployment velocity. Perfect Order Experience now ships features in an afternoon instead of two weeks. WW Grocery cut design document creation from five days to a few hours.
Five steps to becoming a frontier team
The highest-performing teams share five practices with a common logic: reduce the barriers to context for the agent and increase the surface area of work it can do independently. Frontier teams optimize for the rate at which correct, production-ready software reaches customers, not just the speed of individual code generation.
Invest in agent context. The most advanced teams invest heavily in making projects and knowledge easier for agents to consume through agent steering files and guidance on team conventions, coding standards, and testing. The Bedrock infrastructure team placed all code and documentation into a monorepo and kept the inline commentary that AI agents generated, treating it as persistent memory.
Stay updated
Get the day's AI and automation news in your inbox. No spam, unsubscribe anytime.
Slow down to speed up. Every high-performing team reported that things initially slowed down as they learned the models. They encoded cross-functional expertise into reusable steering docs, restructured repositories so LLMs could reason over them, and added comments for AI consumption. The teams that pushed through that learning curve experienced compounding acceleration.
Feed agents instead of babysitting them. Frontier teams maintain a steady backlog of well-scoped tasks with clear outcomes, running multiple agents in parallel and reviewing output asynchronously. Builders report finishing major features in short bursts, with work advancing even when they are not actively waiting.
Make intent explicit before code gets written. Whether through structured specifications, detailed requirements documents, or well-scoped task decomposition, frontier teams ensure agents have clear context about what "done" looks like. Some teams using this approach report handwriting only 1-2% of their code while pushing significantly more commits per person per week.
Shift testing left. Frontier teams build tooling so agents can run all integration tests locally and self-correct before code ever reaches the pipeline. The Prime Video team invested in automated guardrails, component tests, performance tests, and formatters. Code reviews shifted focus to interface definitions and architectural decisions rather than code style and naming.
What technology leaders can do today
Teams that skip the context-building phase, treat AI as a drop-in replacement, or expect immediate gains without restructuring work consistently underperform. Developers across the industry have adopted AI coding tools but not all are seeing production gains. They are using the right tools inside the wrong workflows.
The key takeaways: change how you work to make AI work at its best. Three factors multiply to deliver outcomes: AI handling low-judgment work, uninterrupted focus on high-judgment work, and instant access to domain expertise. The practical starting point is a deliberate pilot with a small team willing to spend the first weeks building agent context before writing production code. Measure commit velocity, deployment frequency, and time-to-resolution. Then use what they learn to build the playbook for the rest of the organization.
The teams achieving 4.5x to more than 10x productivity gains have not just adopted better technology. They have figured out how to work differently with it. That decision is available to every engineering organization today.
This article is based on a blog post by Swami Sivasubramanian, Vice President for Agentic AI at AWS.
Related on Neura Market